The statement you quoted also holds for humans. Of those who can solve grade school math problems, very, very few can solve mathematical problems at a research level.
With recent advancements they can already outperform humans at many tasks that were considered to require AGI level machine intelligence just a few years ago.